Anomalies in demand can be caused due to extraordinary operating conditions such as breakdown of a production system, scheduled turn over, or logistical issues due to extreme weather. To analyze the demand anomalies, firstly exploratory analysis of historical data, spanning at least one to two years, is required. In historical data spanning two or more years, the seasonal trends can be isolated from the overall increase in demand year over year. If there are unusual increase or decline in demand that is not supported by seasonality and growth, then those can be categorized as anomalies. It is also important to have a discussion with the customer based on the trends seen in exploratory data analysis of historical demand and future goals for the business.
Typically, customers have seen at least a few spikes in demand and it is important to understand how the demand spike has affected their overall business in the past. Based on the complexity of the model, it is possible to also understand the reasoning for the unusual spike in demand. Some customers want the model to ignore the anomalies as they represent unusual circumstances that are deceptive when included in the model. Another approach to handling anomalies is to create models with and without anomalies. If the evaluation metrics show that the models without anomalies perform significantly better than the models with anomalies, then the inclusion of the anomalies needs to be reconsidered. If the improvement is not significant, then it is better to include the anomalies and use other modelling techniques to improve the model performance.
Stay tuned for our next blog in this series.
Are you an industrial IoT customer looking to save money by having an efficient demand forecasting model, then talk to us. Our 20 years of experience in IoT combined with our expertise in data science, will help you navigate the complex world of demand and customer behavior. For questions about how Bsquare can help reduce your operational cost, talk to us.